Absolutely! There are definitely ways to help the agent produce different results. My main concern isn’t with the model’s results, but with the “vibe coding” itself. There have been countless stories where people have “vibe coded” an app only for it to be hacked, sensitive keys leaked, or heavy API usage due to lack of optimizations. Using an AI agent isn’t necessarily a bad idea, but realistically, someone who doesn’t understand coding probably won’t be setting up rules or MCP servers either. So, in this example, it was the right decision to not have any rules and just use what came out of the box.
For sure! AI assisted development can be very powerful. It should not, at least in its current states, replace anyone's development jobs. Even scaling back teams due to AI is a mistake unless those teams were already overpopulated from an over-hiring effort. And to your point,... this would not be vibe coding. If we critique and modify the code that is being generated, regardless of how much code it is, we're outside the scope of vibe coding and now in assisted coding. How much assistance the agent provides is a different discussion too. If we rely on the agent to do most/all of the coding and we simply just tweak it, we're going to be heading down a bad path of messy code that will be more expensive to maintain down the road. I don't think that there's a magic percentage of how much generated code is acceptable either. Nor, the implications of where did it get the knowledge to generate that code and who truly owns it.
Absolutely, but I took the approach that has been seen on social media where non-technical people have been using "vibe coding" to create SaaS applications. In that context, they wouldn't know to use TurboStreams and they also may not know the difference between different models if it is producing the results. Claude 3.7 definitely churns out some pretty cool stuff. I haven't looked in to Gemini 2.5 yet.
